Data Augmentation through Generative Models

In recent years, semantic segmentation has emerged as a critical task in computer vision, vital for applications ranging from autonomous driving to medical image analysis. This task involves labeling each pixel in an image with semantic categories such as vehicles, pedestrians, and road signs. However, training deep learning models for semantic segmentation requires substantial amounts of labeled data, which is expensive and time-consuming to acquire. To address this challenge, data augmentation techniques are applied to artificially expand the training set and enhance model performance. Recent studies have highlighted the efficacy of augmenting data in feature space, where extracted features better capture semantic variations compared to raw input data. This approach leads to more realistic augmented data, thereby improving model robustness and generalization. Moreover, GAN-based methods have emerged as promising tools for data augmentation by generating synthetic data that closely resemble real-world data. This thesis investigates GAN-based methods for augmenting data in the feature space for semantic segmentation tasks. The proposed methodology involves extracting features using the encoder of a segmentation model, generating augmented features similar to real features using a GAN or similar generative model, and subsequently training the segmentation model decoder using both real and generated features. Experimental evaluations focus on assessing the effectiveness of this approach in enhancing the accuracy and generalization capabilities of DeepLabV3 segmentation model across Pascal VOC dataset.